Krell Reference 64 digital processor Measurements

Sidebar 3: Measurements

The Reference 64 had a maximum output level of 2.4V from the unbalanced outputs and 4.8V from the balanced jacks—exactly the 6dB increase expected. Channel balance—a measure of how closely matched the left- and right-channel output levels are to each other—was virtually perfect, with the right channel measuring just 0.03dB higher than the left. This excellent performance is no doubt due to the manual gain-trimming at the factory.

Output impedance was a low 16 ohms from the unbalanced outputs and 32 ohms from the balanced outputs, measured at any audio frequency. This low output impedance suggests that the Reference 64 won't interact with the input impedance of preamplifiers, and will drive passive level controls adequately. DC levels at the output were low to moderate, measuring 3.9mV (left channel) and 10.4mV (right) from both the balanced and unbalanced outputs. I heard no pops or ticks through the loudspeakers when I switched between inputs or activated other front-panel functions.

The Reference 64 had no problem locking to 32kHz, 44.1kHz, or 48kHz sampling frequencies. The unit does not invert absolute polarity (a positive-going impulse at the input is positive-going at the output) unless the front-panel "180°" button is pushed. The XLR is wired with pin 2 "hot," meaning the Reference 64 is non-inverting when used with other "pin 2 hot" equipment.

The following measurements were taken from the Reference 64's balanced outputs. If the unbalanced performance was different, it is noted in the text.

First, the Reference 64 had the flattest frequency response of any processor I've measured, being down just 0.04dB at 20kHz (fig.1). This is the result of the custom digital filter—the designers can control the point where the passband ends and the transition band starts. The transition band is the frequency span between where the filter begins rolling off and the point where the attenuation is complete. Most processors are down a few tenths of a dB (0.3dB is typical) at 20kHz. Some software-based processors have much greater treble rolloffs—3dB at 20kHz in one unit. Note in fig.1 how tightly the traces overlap, revealing the perfect level matching between channels.

Fig.1 Krell Reference 64, frequency response and de-emphasis error (bottom) (right channel dashed, 0.5dB/vertical div.).

Also included in fig.1 is the Reference 64's de-emphasis error. There's a slight rise above 7kHz, reaching a maximum error of +0.25dB at 20kHz. This may be barely audible as an increase in upper treble and "air" when playing pre-emphasized discs (of which there are very few). Krell chose to use an analog de-emphasis circuit rather than the DSP chips to perform de-emphasis in the digital domain; the latter would have consumed precious computing cycles.

Fig.2 shows the Reference 64's interchannel crosstalk. It measured –112dB at 1kHz, decreasing to –101dB at 20kHz (Krell specifies 111dB channel separation at 1kHz). This is better-than-average performance, but short of the 130dB (at 1kHz) channel separation seen in the best-measuring processors (footnote 1). Unbalanced crosstalk was identical.

Fig.2 Krell Reference 64, balanced crosstalk (right–left dashed, 10dB/vertical div.).

A spectral analysis of the Reference 64's output when decoding a –90dB, dithered 1kHz sinewave is shown in fig.3. There's a moderate level of noise visible below 500Hz, particularly at the power-line frequency of 60Hz. We can also see some negative linearity error, revealed by the peaks at 1kHz not quite reaching the –90dB horizontal division. Fig.4 is the same type of spectral analysis, but is made with an input signal of all zeros (no signal) and a 200kHz measurement bandwidth. The noise rises gently and inconsequentially above the audio band, with no peaks in the trace that would indicate DAC artifacts.

Fig.3 Krell Reference 64, spectrum of dithered 1kHz tone at –90.31dBFS, with noise and spuriae (1/3-octave analysis, right channel dashed).

Fig.4 Krell Reference 64, spectrum of silent track, 20Hz–200kHz, with noise and spuriae (1/3-octave analysis, right channel dashed).

The Reference 64's linearity, shown in fig.5, was moderately good, with a slight negative error below –75dB that peaked at about –95dB. Interestingly, the linearity improved when I measured the single-ended outputs about an hour after measuring the balanced outputs. Remeasuring the balanced outputs confirmed that the linearity error decreases as the unit warms up. The plot in fig.5 was the best I could get, made after the Reference 64 had been on the bench about 21/2 hours. Even after this warmup time, the Reference 64 didn't get as hot as it normally does when left on continuously. Obviously, the MSB trimmers are adjusted when the unit is at its maximum temperature, and it will perform its best when fully warm. Note that the review sample didn't meet Krell's specification of ±0.3dB linearity error at –90dB. It's possible—though unlikely—that the 2dB and 3.7dB linearity errors at –90dB would be reduced to less than 0.3dB with additional warmup. Incidentally, the Reference 64 takes a very long time to sound its best—it should be left on for at least two days before performing any critical auditioning. Its sound keeps on improving, even after a week!

Fig.5 Krell Reference 64, departure from linearity (right channel dashed, 2dB/vertical div.).

Fig.6 is the Reference 64's reproduction of a 1kHz, undithered sinewave at –90dB. The stairstep waveshape is moderately good, but overlaid with audio-band noise. Note that the horizontal scale is offset (0V to 1mV, rather than the usual ±500µV) because of some DC present.

Fig.6 Krell Reference 64, waveform of undithered 1kHz sinewave at –90.31dBFS.

Much has been made of the squarewave response of custom DSP-based digital filters. Some of them—the Meitner and the Wadia—have minimal overshoot and/or ringing typical of conventional filters. The Reference 64's squarewave response, shown in fig.7, has some overshoot and ringing, but the peaks aren't clipped as they are with the NPC filter chip.

Fig.7 Krell Reference 64, 1kHz squarewave at 0dBFS.

The Reference 64's noise-modulation performance is shown in fig.8. Each trace is the Reference 64's noise floor with a different input level (from –60dB to –100dB; the 41Hz stimulus tone is removed by a high-pass filter). Ideally, the noise floor shouldn't shift as a function of input level—you should see five perfectly overlapping traces. We can see that the Reference 64 has a low noise level, but there is also some divergence of the traces, which indicates that the noise floor increases slightly at low signal levels. The traces, however, remain fairly straight, suggesting that the noise floor's spectral balance doesn't change with input level.

Fig.8 Krell Reference 64, noise modulation, –60 to –100dBFS (10dB/vertical div.).

Fig.9 is an FFT of the Reference 64's output when decoding a full-scale mix of 19kHz and 20kHz sinewaves. (Each component is at –6dBFS, the combined waveform therefore peaks at 0dBFS.) The 1kHz difference component is relatively low in level (–92dB), as are the sidebands around the test-signal frequencies. There is, however, some spurious energy apparent at 3kHz.

Fig.9 Krell Reference 64, HF intermodulation spectrum, DC–22kHz, 19+20kHz at 0dBFS (linear frequency scale, 20dB/vertical div.).

I was unable to measure the Reference 64's word-clock jitter: the DACs are housed in aluminum cases that make them inaccessible unless they're unplugged from the mother board. This is unfortunate, because it would have been possible to measure the effects of the Time Sync function on word-clock jitter. Although I couldn't measure its effect, it was certainly audible.—Robert Harley

Footnote 1: Interestingly, the best phono cartridges manage only about 30dB of channel separation at 1kHz—on a good day.—Robert Harley
Krell Industries
45 Connair Road
Orange, CT 06477-3650
(203) 298-4010